Beyond Accuracy: A Stability‐Focused Framework for Raman Spectroscopic Margin Assessment

ABSTRACT High classification accuracy alone does not guarantee reliable intraoperative margin assessment, particularly for Raman spectroscopic diagnostics that are sensitive to spatial heterogeneity and sampling variability. This study systematically examines how spectral averaging, spatial sampling design, and training cohort composition affect the stability and generalizability of Raman mapping‐based tissue classification for tongue squamous cell carcinoma. Raman mapping data were acquired from paired tumor and adjacent normal tissues, followed by two‐dimensional spatial averaging across multiple integration scales. A range of conventional machine learning and deep learning models were evaluated, with XGBoost and ResNet selected for controlled comparisons. Model performance was assessed under varying window sizes, sampling strategies, and training cohort compositions using repeated cross‐validation. Increasing the spatial averaging scale led to a systematic improvement in discrimination performance, with mean AUC rising rapidly at small window sizes and approaching saturation at larger scales, while performance variability showed no systematic dependence on window size. Interleaved spatial sampling consistently improved robustness to intratissue heterogeneity. Expanding the training cohort from five to 10 patients markedly enhanced cross‐patient generalization, with no repeated trial falling below a clinically relevant threshold (AUC ≥ 0.90). While both models achieved comparable average accuracy, XGBoost exhibited greater stability under limited training data. Rather than introducing new classification architectures, this work defines quantitative parameter regimes that balance accuracy, stability, and feasibility, providing a stability‐oriented framework to support the clinical translation of Raman spectroscopic margin assessment.